diff --git a/dev-util/nvidia-cuda-toolkit/Manifest b/dev-util/nvidia-cuda-toolkit/Manifest new file mode 100644 index 0000000..29279f8 --- /dev/null +++ b/dev-util/nvidia-cuda-toolkit/Manifest @@ -0,0 +1,7 @@ +AUX cuda-config.in 407 BLAKE2B 5f7a0d50d774db58c180b767f00bead7c7dbf5f6b3b4a2f960dfc3b8450f7abf5c2b67c6620c5318f1331646dd2003d61c70a4c31359efd16ff418ecc9d0b446 SHA512 363903b999c43a47a48c52e8898c6044fadcc66d259338ee22863e859c2bc9873a2f1392aa83c9182486d3aa4eaeb5f164cf8ccf29cc74fd8ad56245f76a039a +AUX nvidia-cuda-toolkit-glibc-2.41-r1.patch 4855 BLAKE2B e06dc68c473fb427c8347343652dbd93f209fb0b19ade54aba7bb2a6ab5344b896d1f2dc86ed4e0ec1a4de739716fe30ff1ff3629d73eb6a4c081c25a7454fae SHA512 037c906903cf7058543c6636263d93ce5b80bedcc601417c7e23758966c428434593d61ee7e32d51e07c5f16ee595510f1ce7db9f000f906fe96928a3b4ed5fb +AUX parse_manifest.py 8488 BLAKE2B 8ac93a39276eaf89b41a5b0baa168b5974423077b0fd4b22841451514513c914fb4188863d73591f89975cc833dcb88bde289b0676f218715d8735a2d5c0f159 SHA512 4b2d9de215858d25319254666a22a184e26510a3488fc97a939704fab1e81a9baa363fc1168129618ce076de84d7acbeb71d98b8be272017f47b54c237d9b9a4 +DIST cuda_12.9.1_575.57.08_linux.run 5860276058 BLAKE2B 482aee91731ca490956f3cf375e973b15ddcf69b295693d34488958e2592e5c55dabed589aa4f5f4ad59a50ae00b4523e5ebf760a613b634ca00164d363fb567 SHA512 fc29c5fc1121fb6634f1fe396abe7f34d351686454516269e9143e678ea178f906a35b916b8bb2d96ecfcfc705dda7d0f4547f7e7f00d36e392d981a766b6a56 +DIST cuda_12.9.1_575.57.08_linux_sbsa.run 5074335372 BLAKE2B 71782c3e7e39503aa914c8e02b7e5e44c856bd98bb9d1d46b90dfb7db4a2becb03ed1467fbca34c2ea1147202deb379198495007d69c7987320e138a63f2164e SHA512 91e3b457b6c36c8c3f07ae6fbd24f3f90b9ea22397059f91655ad68c32b96547e2c732af987a0c5ea63c37c43d48f199d97b9e84068c145633ebf67c34465ac6 +EBUILD nvidia-cuda-toolkit-12.9.1.ebuild 10590 BLAKE2B bd84a00b2c334de9625a6d00250372b3a227746d22efec9f24f6679523a45cbd39825bc28fb002abced5fa447769d416c0706362cf3d549074a2d42e46520c44 SHA512 4cafc7bdbc64d039cd922286ddb990dc8544a23427191c639bc36bbd93117521d7e49efb363a2a448158f9b392f2357cf41c82785dff5475ec07985b78ebd6fd +MISC metadata.xml 843 BLAKE2B 5ed06252572415290175ff90b66c1e13b67d282085295586e046019cc7d28b6e11728fba310bcf4101caaf09f4d1383f8df3f1c97325fa9e979dfa06e4704571 SHA512 208c83e0d968796d526b7223c5221f0805401b31420807a21f078797cba8e3f141c654017331a89649af2aa29a40cc1ff28f9cfa3ba2e6bee85fa76957c01837 diff --git a/dev-util/nvidia-cuda-toolkit/files/cuda-config.in b/dev-util/nvidia-cuda-toolkit/files/cuda-config.in new file mode 100644 index 0000000..10db7dd --- /dev/null +++ b/dev-util/nvidia-cuda-toolkit/files/cuda-config.in @@ -0,0 +1,28 @@ +#!/bin/bash + +SUPPORT_GCC_VERSIONS_BY_CUDA="CUDA_SUPPORTED_GCC" + +_print_help() { + cat <<- EOF + Usage: + $(basename $0) [options] + + -s | --supported Returns by current CUDA supported gcc versions + -h | --help Shows this help + EOF +} + +case ${1} in + -s|--supported) + echo "${SUPPORT_GCC_VERSIONS_BY_CUDA}" + exit 0 + ;; + -h|--help) + _print_help + exit 255 + ;; + *) + _print_help + exit 1 + ;; +esac diff --git a/dev-util/nvidia-cuda-toolkit/files/nvidia-cuda-toolkit-glibc-2.41-r1.patch b/dev-util/nvidia-cuda-toolkit/files/nvidia-cuda-toolkit-glibc-2.41-r1.patch new file mode 100644 index 0000000..32e5bd6 --- /dev/null +++ b/dev-util/nvidia-cuda-toolkit/files/nvidia-cuda-toolkit-glibc-2.41-r1.patch @@ -0,0 +1,119 @@ +From cae8ab12c9f981f110bb2e1318d9c5306e3a2d81 Mon Sep 17 00:00:00 2001 +From: Paul Zander +Date: Sat, 8 Feb 2025 16:34:42 +0100 +Subject: [PATCH] Updates headers for glibc-2.41 + +Match the extern definitions in cuda-crt headers with the changes in glibc-2.41. + +/usr/include/bits/mathcalls.h(79): error: exception specification is +incompatible with that of previous function "cospi" (declared at line 2601 +of +/opt/cuda-12.8.0/bin/../targets/x86_64-linux/include/crt/math_functions.h) + + extern double cospi (double __x) noexcept (true); extern double __cospi (double __x) noexcept (true); + +/opt/cuda-12.8.1/bin/../targets/x86_64-linux/include/crt/math_functions.hpp:2982:17: error: 'sinpi' is missing exception specification 'throw()' + 2982 | __func__(double sinpi(double a)) + | ^ + | throw() +/opt/cuda-12.8.1/bin/../targets/x86_64-linux/include/crt/func_macro.h:37:16: note: expanded from macro '__func__' + 37 | inline decl + | ^ +/opt/cuda-12.8.1/bin/../targets/x86_64-linux/include/crt/math_functions.h:6006:17: note: previous declaration is here + 6006 | __func__(double sinpi(double a)); + +See-Also: https://sourceware.org/git/?p=glibc.git;a=commit;h=0ae0af68d8fa3bf6cbe1e4f1de5929ff71de67b3 +Signed-off-by: Paul Zander + +diff --git a/builds/cuda_nvcc/targets/x86_64-linux/include/crt/math_functions.h b/builds/cuda_nvcc/targets/x86_64-linux/include/crt/math_functions.h +index d8201f9..a9b19d4 100644 +--- a/builds/cuda_nvcc/targets/x86_64-linux/include/crt/math_functions.h ++++ b/builds/cuda_nvcc/targets/x86_64-linux/include/crt/math_functions.h +@@ -2553,7 +2553,11 @@ extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float rcbrt + * + * \note_accuracy_double + */ ++#if defined(__GLIBC__) && __GLIBC__ == 2 && __GLIBC_MINOR__ >= 41 ++extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double sinpi(double x) noexcept (true); ++#else + extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double sinpi(double x); ++#endif + /** + * \ingroup CUDA_MATH_SINGLE + * \brief Calculate the sine of the input argument +@@ -2576,7 +2580,11 @@ extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double sinpi + * + * \note_accuracy_single + */ ++#if defined(__GLIBC__) && __GLIBC__ == 2 && __GLIBC_MINOR__ >= 41 ++extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float sinpif(float x) noexcept (true); ++#else + extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float sinpif(float x); ++#endif + /** + * \ingroup CUDA_MATH_DOUBLE + * \brief Calculate the cosine of the input argument +@@ -2598,7 +2606,11 @@ extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float sinpi + * + * \note_accuracy_double + */ ++#if defined(__GLIBC__) && __GLIBC__ == 2 && __GLIBC_MINOR__ >= 41 ++extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double cospi(double x) noexcept (true); ++#else + extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double cospi(double x); ++#endif + /** + * \ingroup CUDA_MATH_SINGLE + * \brief Calculate the cosine of the input argument +@@ -2620,7 +2632,11 @@ extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double cospi + * + * \note_accuracy_single + */ ++#if defined(__GLIBC__) && __GLIBC__ == 2 && __GLIBC_MINOR__ >= 41 ++extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float cospif(float x) noexcept (true); ++#else + extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float cospif(float x); ++#endif + /** + * \ingroup CUDA_MATH_DOUBLE + * \brief Calculate the sine and cosine of the first input argument +diff --git a/builds/cuda_nvcc/targets/x86_64-linux/include/crt/math_functions.hpp b/builds/cuda_nvcc/targets/x86_64-linux/include/crt/math_functions.hpp +index cc09b91..11eec71 100644 +--- a/builds/cuda_nvcc/targets/x86_64-linux/include/crt/math_functions.hpp ++++ b/builds/cuda_nvcc/targets/x86_64-linux/include/crt/math_functions.hpp +@@ -2979,7 +2979,7 @@ __func__(double rcbrt(const double a)) + return t; + } + +-__func__(double sinpi(double a)) ++__func__(double sinpi(double a)) throw() + { + int n; + +@@ -3009,7 +3009,7 @@ __func__(double sinpi(double a)) + return a; + } + +-__func__(double cospi(double a)) ++__func__(double cospi(double a)) throw() + { + int n; + +@@ -3340,12 +3340,12 @@ __func__(float rcbrtf(const float a)) + return static_cast(rcbrt(static_cast(a))); + } + +-__func__(float sinpif(const float a)) ++__func__(float sinpif(const float a)) throw() + { + return static_cast(sinpi(static_cast(a))); + } + +-__func__(float cospif(const float a)) ++__func__(float cospif(const float a)) throw() + { + return static_cast(cospi(static_cast(a))); + } +-- +2.48.1 + diff --git a/dev-util/nvidia-cuda-toolkit/files/parse_manifest.py b/dev-util/nvidia-cuda-toolkit/files/parse_manifest.py new file mode 100644 index 0000000..d50545c --- /dev/null +++ b/dev-util/nvidia-cuda-toolkit/files/parse_manifest.py @@ -0,0 +1,223 @@ +#!/usr/bin/env python +# Copyright 2024-2025 Gentoo Authors +# Distributed under the terms of the GNU General Public License v2 +# +# +# Takes the manifests/*.xml file from the NVIDIA CUDA Toolkit as input. +# Those files contain a nested tree of package items. +# +# For each package node a if block is output that checks the package's +# name attribute against the environment provided array SKIP_COMPONENTS +# and skips that whole package if it's listed in it. +# +# Each item contains four nodes that reference things to install, +# `dir`, `file`, `desktopFile`, & `pcfile`, and metadata that further +# detail these. This script will output calls to do* calls that +# use the metadata. +# - dodir would create an empty file - we handle this in dofile +# - dofile copies a regex file glob +# - dodesktopFile creates a .desktop file +# - dopcfile creates a pkgconfig file +# +# The resulting bash code can be run inside src_install(). +# +# Usage: python parse_manifest.py + + +import argparse +import xml.etree.ElementTree + +import defusedxml.ElementTree +from pathlib import Path + +ind = "\t" +indent = 1 + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument('filename') # positional argument + + args = parser.parse_args() + + basedir = Path(args.filename).parents[1] + + et = defusedxml.ElementTree.parse(args.filename) + + # Find all packages under the package with the id main (i.e. under "CUDA Installer") + for e in et.findall("[@id='main']/package"): + def p_package(el: xml.etree.ElementTree.Element, level: int = 0): + + skip = { + "Documentation", # obsolete + "Driver", # unused + # "Kernel Objects", # split + # "Demo Suite", + # "Visual Tools", + # old eclipse + # "nsight", + # old java + # "nvvp", + # "cuda-gdb-src" + } + + name = el.get("name") + + # trim leading CUDA and trailing version + if name.startswith("CUDA"): + name2 = ' '.join(name.split(" ")[1:-1]) + else: + name2 = name + + if name2 in skip: + return + + # avoid having to deal with whitespaces in bash + name2 = name2.replace(" ", "_") + + path = "" + + print(f"{ind * (level + 0) * indent}if ! has {name2} \"${{SKIP_COMPONENTS[@]}}\"; then # \"{name}\"") + + # output attributes from unhandled tags + for child in el: + if child.tag == "package": + continue + if child.tag == "file": + continue + if child.tag == "desktopFile": + continue + if child.tag == "pcfile": + continue + for attrib in child.attrib: + print(f"{child.tag} {attrib}={child.attrib[attrib]}") + + # CUDA Installer + # only for CUDA Installer + # unused + for node in el.findall("./name"): + print(f"{ind * (level + 1) * indent}# {node.tag}: \"{node.text}\"") + + # compiler + # category. We use the package name instead. + # unused + # for node in el.findall("./type"): + # print(f"{ind * (level + 1) * indent}# {node.tag}: \"{node.text}\"") + + # 1 + # probably sorting for the tui installer + # unused + # for node in el.findall("./priority"): + # print(f"{ind * (level + 1) * indent}# {node.tag}: \"{node.text}\"") + + # + # proably for tui installer + # unused + # for node in el.findall("./single-selection"): + # print(f"{ind * (level + 1) * indent}# {node.tag}: \"{node.text}\"") + + # 2.24.2 + # version of the installed kernel object (Kernel Objects only) + # unused + for node in el.findall("./koversion"): + print(f"{ind * (level + 1) * indent}# {node.tag}: \"{node.text}\"") + + # /usr/local/cuda-12.8 + # overrides the install location + # unused + # for node in el.findall("./installPath"): + # print(f"{ind * (level + 1) * indent}# {node.tag}: \"{node.text}\"") + + # ./builds/cuda_cccl/ + # path where the package files are found + # we cd into it + for node in el.findall("./buildPath"): + path = node.text.removeprefix('./') + print(f"{ind * (level + 1) * indent}cd \"${{S}}/{path}\" || die \"cd ${{S}}/{path} failed\"") + print() + + # bin + # would install empty dirs + # unused + # for node in el.findall("./dir"): + # pass + + # .* + # targets/x86_64-linux/lib/.*\.so + # regex glob of files to install. + for node in el.findall("./file"): + # unescape '.*' -> '*' & '\.' -> '.' + file = (node.text + .replace(".*", "*") + .replace(r"\.", ".") + .replace("x86_64", "${narch}") + .replace("sbsa", "${narch}") + ) + + # optional dir offset, we merge it into path + dir = "" + if "dir" in node.attrib: + dir = f" \"{Path(node.attrib['dir'])}\"" + + filepath = basedir / path / file + + # ignore existing symlinks ( include, lib* ) and the uninstallers + if not filepath.is_symlink() and not file.endswith("-uninstaller"): + print(f"{ind * (level + 1) * indent}dofile \"{file}\"{dir}") + + # opencl-12.8.pc + # create a pkgconfig file for the given description and lib name/version + for node in el.findall("./pcfile"): + offset = node.text.rfind('-') + if offset == -1: + raise RuntimeError(f"failed to split pcfile {node.text}") + + lib_name = node.text[:offset] + + if not node.text.endswith('.pc'): + raise RuntimeError(f"pcfile does not end in '.pc' {node.text}") + lib_version = node.text[offset+1:-3] + + if "description" not in node.attrib: + raise RuntimeError(f"no description for {node.text}") + + subdir = "" + if "subdir" in node.attrib: + subdir = f" \"{node.attrib['subdir']}\"" + + print(f"{ind * (level + 1) * indent}dopcfile " + f"\"{lib_name}\" " + f"\"{lib_version}\" " + f"\"{node.attrib['description']}\"{subdir}") + + # + # create a .desktop file + for node in el.findall("./desktopFile"): + print(f"{ind * (level + 1) * indent}dodesktopFile \\") + print(f"{ind * (level + 2) * indent}\"{node.attrib['filename']}\" \\") + print(f"{ind * (level + 2) * indent}\"{node.attrib['name']}\" \\") + print(f"{ind * (level + 2) * indent}\"{node.attrib['categories']}\" \\") + print(f"{ind * (level + 2) * indent}\"{node.attrib['keywords']}\" \\") + print(f"{ind * (level + 2) * indent}\"{node.attrib['iconPath']}\" \\") + print(f"{ind * (level + 2) * indent}\"{node.attrib['execPath']}\" \\") + print(f"{ind * (level + 2) * indent}\"{node.attrib['tryExecPath']}\"") + + # iterator over all nested packages + for node in el.findall("./package"): + p_package(node, level + 1) + + print(f"{ind * (level + 0) * indent}fi") + + p_package(e) + + +if __name__ == "__main__": + main() diff --git a/dev-util/nvidia-cuda-toolkit/metadata.xml b/dev-util/nvidia-cuda-toolkit/metadata.xml new file mode 100644 index 0000000..b28664c --- /dev/null +++ b/dev-util/nvidia-cuda-toolkit/metadata.xml @@ -0,0 +1,17 @@ + + + + + sci@gentoo.org + Gentoo Science Project + + + Depend on llvm-core/clang as host compiler + Install the CUDA debugger + Install profiling and optimizing tools (nsight-compute, nsight-systems) + Install the NVIDIA CUDA profiler (nvprof) and the related libraries + Enable infiniband support via sys-cluster/rdma-core + Install compute-sanitizer tool + Install the NVIDIA CUDA visual profiler (nvvp) + + diff --git a/dev-util/nvidia-cuda-toolkit/nvidia-cuda-toolkit-12.9.1.ebuild b/dev-util/nvidia-cuda-toolkit/nvidia-cuda-toolkit-12.9.1.ebuild new file mode 100644 index 0000000..5384ec1 --- /dev/null +++ b/dev-util/nvidia-cuda-toolkit/nvidia-cuda-toolkit-12.9.1.ebuild @@ -0,0 +1,382 @@ +# Copyright 1999-2025 Gentoo Authors +# Distributed under the terms of the GNU General Public License v2 + +# shellcheck disable=SC2317 + +EAPI=8 + +PYTHON_COMPAT=( python3_{11..13} ) +inherit check-reqs edo toolchain-funcs +inherit python-r1 + +DRIVER_PV="575.57.08" +GCC_MAX_VER="14" +CLANG_MAX_VER="19" + +DESCRIPTION="NVIDIA CUDA Toolkit (compiler and friends)" +HOMEPAGE="https://developer.nvidia.com/cuda-zone" +SRC_URI=" + amd64? ( + https://developer.download.nvidia.com/compute/cuda/${PV}/local_installers/cuda_${PV}_${DRIVER_PV}_linux.run + ) + arm64? ( + https://developer.download.nvidia.com/compute/cuda/${PV}/local_installers/cuda_${PV}_${DRIVER_PV}_linux_sbsa.run + ) +" +S="${WORKDIR}" + +LICENSE="NVIDIA-CUDA" + +SLOT="0/${PV}" # UNSLOTTED +# SLOT="${PV}" # SLOTTED + +KEYWORDS="-* ~amd64 ~arm64 ~amd64-linux ~arm64-linux" +IUSE="clang debugger examples nsight profiler rdma sanitizer" +RESTRICT="bindist mirror strip test" + +REQUIRED_USE="${PYTHON_REQUIRED_USE}" + +# since CUDA 11, the bundled toolkit driver (== ${DRIVER_PV}) and the +# actual required minimum driver version are different. +RDEPEND=" + !clang? ( + /dev/null || die + eapply -p5 "${FILESDIR}/nvidia-cuda-toolkit-glibc-2.41-r1.patch" + popd >/dev/null || die + + default +} + +src_configure() { + : +} + +src_compile() { + : +} + +src_install() { + local -x SKIP_COMPONENTS=( + "Kernel_Objects" + "Visual_Tools" + "Documentation" # obsolete + "cuda-gdb-src" # not used + ) + + ! use debugger && SKIP_COMPONENTS+=( "cuda-gdb" ) + ! use examples && SKIP_COMPONENTS+=( "Demo_Suite" ) + ! use profiler && SKIP_COMPONENTS+=( "cuda-cupti" "cuda-profiler-api" "nvprof" ) + ! use sanitizer && SKIP_COMPONENTS+=( "compute-sanitizer" ) + + dodir "${CUDA_PATH}" + into "${CUDA_PATH}" + + dofile() { + debug-print-function "${FUNCNAME[0]}" "$@" + + if [[ $# -ne 1 ]] && [[ $# -ne 2 ]]; then + die "${FUNCNAME[0]} must receive one or two arguments" + fi + + local _DESTDIR + _DESTDIR="$(dirname "${CUDA_PATH}/${1%/}")${2:+/${2%/}}" + mkdir -p "${ED}${_DESTDIR}" || die "mkdir ${_DESTDIR} failed" + + local dir + dir="$(dirname "${1}")" + + if test -z "$(find "${dir}" -maxdepth 1 -name "$(basename "$1")" -print -quit)"; then + if [[ -e "${ED}${_DESTDIR}/$(basename "${1}")" ]]; then + return + fi + if [[ "$1" == "targets/x86_64-linux/lib/stubs/libcusolverMg*" ]] || + [[ "$1" == "targets/x86_64-linux/lib/libcusparse.so.*" ]]; then + return + fi + return + fi + + if [[ -e "${ED}${_DESTDIR}/$(basename "${1}")" ]]; then + # skip noisy warnings + if [[ "$(basename "${1}")" == "include" ]] || + [[ "$(basename "${1}")" == "lib64" ]]; then + return + fi + eqawarn "${ED}${_DESTDIR}/$(basename "${1}") exists" + return + fi + + eval mv -i "${1}" "${ED}${_DESTDIR}" || die "mv failed ${PWD} / ${1} -> ${ED} ${_DESTDIR}" + } + + dopcfile() { + [[ $# -eq 0 ]] && return + + dodir "${CUDA_PATH}/pkgconfig" + cat > "${ED}${CUDA_PATH}/pkgconfig/${1}.pc" <<-EOF || die "dopcfile" + cudaroot=${EPREFIX}${CUDA_PATH} + libdir=\${cudaroot}/targets/${narch}-linux/lib${4} + includedir=\${cudaroot}/targets/${narch}-linux/include + + Name: ${1} + Description: ${3} + Version: ${2} + Libs: -L\${libdir} -l${1} + Cflags: -I\${includedir} + EOF + } + + pushd builds >/dev/null || die + fix_executable_bit=( + cuda_cupti/extras/CUPTI/samples/pc_sampling_utility/pc_sampling_utility_helper.h + cuda_cupti/extras/CUPTI/samples/pc_sampling_continuous/libpc_sampling_continuous.pl + + libcufile/gds/tools/run_gdsio.cfg + ) + + if use amd64; then + fix_executable_bit+=( + cuda_opencl/targets/*/include/CL/cl.hpp + + libcufile/targets/*/lib/libcufile_rdma_static.a + libcufile/targets/*/lib/libcufile_static.a + ) + fi + chmod -x "${fix_executable_bit[@]}" || die "failed chmod" + popd >/dev/null || die + + ebegin "parsing manifest" "${S}/manifests/cuda_"*.xml # {{{ + + "${EPYTHON}" "${FILESDIR}/parse_manifest.py" "${S}/manifests/cuda_"*".xml" &> "${T}/install.sh" \ + || die "failed to parse manifest" + # shellcheck disable=SC1091 + source "${T}/install.sh" || die "failed to source install script" + + eend $? # }}} + + if use debugger; then + if [[ -d "${ED}/${CUDA_PATH}/extras/Debugger/lib64" ]]; then + rmdir "${ED}/${CUDA_PATH}/extras/Debugger/lib64" || die "remove debugger lib64" + fi + + find "${ED}/${CUDA_PATH}/bin" -maxdepth 1 -name "cuda-gdb-*-tui" -print0 | while read -rd $'\0' tui_name; do + impl="$(basename "${tui_name}" | cut -d '-' -f 3 | tr '.' '_')" + + if ! has "${impl}" "${PYTHON_COMPAT[@]}" || ! use "python_targets_${impl}"; then + rm "${tui_name}" || die "tui-name rm ${tui_name}" + sed -e "/$(basename "${tui_name}")\"/d" -i "${ED}/${CUDA_PATH}/bin/cuda-gdb" || die "tui_name sed" + fi + done + fi + + # remove rdma libs (unless USE=rdma) + if ! use rdma; then + rm "${ED}/${CUDA_PATH}/targets/${narch}-linux/lib/libcufile_rdma"* || die "failed to remove rdma files" + fi + + # Add include and lib symlinks + dosym -r "${CUDA_PATH}/targets/${narch}-linux/include" "${CUDA_PATH}/include" + dosym -r "${CUDA_PATH}/targets/${narch}-linux/lib" "${CUDA_PATH}/$(get_libdir)" + + find "${ED}/${CUDA_PATH}" -empty -delete || die + + local ldpathextradirs pathextradirs + + use debugger && ldpathextradirs+=":${EPREFIX}${CUDA_PATH}/extras/Debugger/lib64" + use profiler && ldpathextradirs+=":${EPREFIX}${CUDA_PATH}/extras/CUPTI/lib64" + + local revord=$(( 999999 - $(printf "%02d%02d%02d" "$(ver_cut 1)" "$(ver_cut 2)" "$(ver_cut 3)") )) + newenvd - "99cuda${revord}" <<-EOF + PATH=${EPREFIX}${CUDA_PATH}/bin${pathextradirs} + PKG_CONFIG_PATH=${EPREFIX}${CUDA_PATH}/pkgconfig + LDPATH=${EPREFIX}${CUDA_PATH}/$(get_libdir):${EPREFIX}${CUDA_PATH}/nvvm/lib64${ldpathextradirs} + EOF + + # CUDA prepackages libraries, don't revdep-build on them + insinto /etc/revdep-rebuild + newins - "80${PN}${revord}" <<-EOF + SEARCH_DIRS_MASK="${EPREFIX}${CUDA_PATH}" + EOF + + # https://bugs.gentoo.org/926116 + insinto /etc/sandbox.d + newins - "80${PN}" <<-EOF + SANDBOX_PREDICT="/proc/self/task" + EOF + + # TODO drop and replace with runtime detection similar to what python does {{{ + # ATTENTION: change requires revbump, see link below for supported GCC # versions + # https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#system-requirements + local cuda_supported_gcc=( 8.5 9.5 10 11 12 13 "${GCC_MAX_VER}" ) + + sed \ + -e "s:CUDA_SUPPORTED_GCC:${cuda_supported_gcc[*]}:g" \ + "${FILESDIR}"/cuda-config.in > "${ED}/${CUDA_PATH}/bin/cuda-config" || die + fperms +x "${CUDA_PATH}/bin/cuda-config" + # }}} + + # skip til cudnn has been changed #950207 + # if [[ "${SLOT}" != "${PV}" ]]; then + # dosym -r "${CUDA_PATH}" "${CUDA_PATH%"-${PV}"}" + # fi + + fowners -R root:root "${CUDA_PATH}" +} + +pkg_postinst_check() { + if tc-is-gcc || ! use clang; then + if ver_test "$(gcc-major-version)" -gt "${GCC_MAX_VER}"; then + ewarn + ewarn "gcc > ${GCC_MAX_VER} will not work with CUDA" + ewarn + ewarn "Append --ccbin= pointing to a gcc bindir to the nvcc compiler flags (NVCCFLAGS)" + ewarn "or set NVCC_CCBIN to the same bindir." + ewarn " NVCCFLAGS=\"--ccbin=$(eval echo "${EPREFIX}/usr/*-linux-gnu/gcc-bin/${GCC_MAX_VER}")\"" + ewarn " NVCC_CCBIN=$(eval echo "${EPREFIX}/usr/*-linux-gnu/gcc-bin/${GCC_MAX_VER}")" + ewarn + fi + fi + + if tc-is-clang || use clang; then + if ver_test "$(clang-major-version)" -gt "${CLANG_MAX_VER}"; then + ewarn + ewarn "clang > ${CLANG_MAX_VER} will not work with CUDA" + ewarn + ewarn "Append --ccbin= pointing to a clang bindir to the nvcc compiler flags (NVCCFLAGS)" + ewarn "or set NVCC_CCBIN to the same bindir." + ewarn " NVCCFLAGS=\"--ccbin=$(eval echo "${EPREFIX}/usr/lib/llvm/*/bin${CLANG_MAX_VER}")\"" + ewarn " NVCC_CCBIN=$(eval echo "${EPREFIX}/usr/lib/llvm/*/bin${CLANG_MAX_VER}")" + ewarn + fi + fi +} + +pkg_postinst() { + if [[ ${MERGE_TYPE} != binary ]]; then + pkg_postinst_check + fi + + if use profiler; then + einfo + einfo "nvidia-drivers restricts access to performance counters." + einfo "You'll need to run profiling tools (nvprof) " + einfo "using sudo (needs cap SYS_ADMIN) or add the following line to " + einfo "a modprobe configuration file " + einfo "(e.g. /etc/modprobe.d/nvidia-prof.conf): " + einfo + einfo "options nvidia NVreg_RestrictProfilingToAdminUsers=0" + einfo + fi +} diff --git a/metadata/md5-cache/dev-util/nvidia-cuda-toolkit-12.9.1 b/metadata/md5-cache/dev-util/nvidia-cuda-toolkit-12.9.1 new file mode 100644 index 0000000..6c30e3c --- /dev/null +++ b/metadata/md5-cache/dev-util/nvidia-cuda-toolkit-12.9.1 @@ -0,0 +1,16 @@ +BDEPEND=|| ( ( dev-lang/python:3.13 dev-python/defusedxml[python_targets_python3_13(-)] ) ( dev-lang/python:3.12 dev-python/defusedxml[python_targets_python3_12(-)] ) ( dev-lang/python:3.11 dev-python/defusedxml[python_targets_python3_11(-)] ) ) +DEFINED_PHASES=compile configure install postinst prepare pretend setup unpack +DESCRIPTION=NVIDIA CUDA Toolkit (compiler and friends) +EAPI=8 +HOMEPAGE=https://developer.nvidia.com/cuda-zone +INHERIT=check-reqs edo toolchain-funcs python-r1 +IUSE=clang debugger examples nsight profiler rdma sanitizer python_targets_python3_11 python_targets_python3_12 python_targets_python3_13 +KEYWORDS=-* ~amd64 ~arm64 ~amd64-linux ~arm64-linux +LICENSE=NVIDIA-CUDA +RDEPEND=!clang? (